Two novel composite kernels for relation extraction

Relation extraction aims at discovering relations between entities from free text, and it is a crucial part of information extraction. Recently, kernel methods have seen successfully applied in relation extraction. The paper proposes two novel composite kernels for relation extraction, namely linear and polynomial kernels, based on three individual kernels: an entity kernel that allows for structured features, a string kernel for parse tree, and Zelenko´s parse tree kernel. In experiments, the kernels mentioned above are used in conjunction with Support Vector Machines for extracting person-affiliation relations from 500 sentences. In order to improve the training speed, trees parsed from Stanford Parser are pruned before using. Finally, the outcome shows that though linear composite kernel´s precision (77.0%) and recall (82.2%) are not the highest, its F-measure with 79.4% significantly outperforms the best record, which is 72.6% of three previous kernels. This result indicates that the linear composite kernel performs better than the three individual kernels.